The number everyone quotes is around two billion dollars per approved drug, and the usual gloss is that discovery is hard. Discovery is hard, but that isn’t where the two billion comes from. It comes from everything that fails on the way to the one thing that doesn’t.
The model below makes that concrete. It follows a cohort of 100 single-asset biotech companies from pre-clinical work through FDA review, drawn as two Sankey diagrams on a calendar timeline. Most of them fail, and it tracks exactly where and when: some run out of money, most fail the science, and only a handful reach approval. Press Run both and watch the cohort resolve.
Both diagrams run the identical cohort with identical random draws — the same molecule gets the same luck in each — so any difference between them is caused by the switches, not by chance. Left to right is years.
| Compare | ■ Scenario A | ■ Scenario B | B vs A |
|---|---|---|---|
| Cost per approved drug | $977M | $529M | -46% |
| Approvals per cohort | 2.1 | 4.2 | +95% |
| Capital consumed | $2.10B | $2.21B | +5% |
| Viable drugs lost to cash would have been approved with money | 1.8 | 3.9 | +123% |
| Pipeline length | 13.7 yrs | 13.7 yrs | 0% |
Both bars are cut by phase, on the same colours as the diagrams below: the live ramp for money still in play, the rust ramp for money lost at that phase. Segments run pre-clinical to review, left to right, so the written-off bar reddens toward the right as the losses move later and get more expensive.
Both diagrams share the same pixels-per-year and the same vertical scale, so lengths and thicknesses are directly comparable — a scenario that finishes sooner is drawn shorter. The green and rust figures beside each lane are that lane's difference against the other scenario. Blue is still in play — pale inside a stage, solid once past it. Rust is failed, black is out of money, green is approved.
Long-run average — specialists, selective. Sets the odds of closing each round — and so the size of the black lane and how many viable drugs die of money rather than science.
The full gate: a market-driven raise curve, a penalty on programmes that drag, and easier raises for programmes that have compressed their time and cost.
Each switch flipped on by itself, measured two ways: its effect on cost per approved drug and on PTS — the probability of technical success, i.e. the odds a programme clears every scientific gate. The split is the whole point: time-and-money levers move cost and leave PTS untouched, while odds levers move both.
Baseline at these settings: $977M per approved drug, 4.1% PTS. Bars are scaled within each column. The ordering by cost impact holds across markets, but the magnitudes move — odds levers pull further ahead of time levers as the market worsens, because a higher PTS also rescues more drugs from dying of cash.
Generative chemistry & in-silico screening — Model-designed candidates and computational triage before the bench. Cheaper, faster candidates — who then face exactly the same clinical gauntlet.
AI site selection & activation — Ranking investigator sites on predicted enrollment instead of prior relationships. Vendor-reported, unaudited.
AI patient matching & enrollment — Screening records against protocol criteria to fill cohorts faster. The biggest cycle-time claim in the vendor literature, and the least audited.
Automated submission drafting — Generating CSRs and module text from trial data. Real savings, on the shortest and cheapest stage in the pipeline — watch how little it moves.
Run Phase I in China — First-in-human in China: direct costs about 60% lower and enrollment two to three times faster, on a large pool of treatment-naive patients. Phase II onward stays in the US, so nothing here turns on the single-country data question.
In-silico ADMET & toxicity prediction — Predicting absorption, metabolism and tox before dosing humans. Jayatunga et al. 2024 found AI-discovered molecules clear Phase I at 80–90% — real, and the cheapest phase to win.
Genetically supported target selection — Choosing targets with human genetic causal evidence, which ML is what makes tractable at genome scale. Nelson et al. 2015 estimated a doubling of clinical success; Minikel/Nelson 2024 refined it to 2.6x; King et al. replicated it and located the effect in Phases II and III. Only about 15% of programmes currently have that support — the best-evidenced lever here by some distance.
The area you pick changes every transition rate, duration and cost in the model. Here is what is behind those numbers.
One asset per company. The unit here is a company whose fate is one molecule's fate — it lives or dies with its lead program. That is deliberately not the whole industry: in the US, platform companies outnumber single-asset ones by roughly three to one among venture-backed biotechs, because larger funds prefer bets where one failure doesn't end the company. But single-asset and single-lead-asset companies remain common, especially at the clinical-stage IPOs that dominate the funnel — and even nominal platforms are often, in practice, valued and funded on one lead program. The single-asset frame is chosen because a molecule's journey maps cleanly onto a company's; it overstates binary company death relative to a diversified platform, which reallocates rather than dies when a program fails.
The financing gate, with its actual numbers. A round is required to begin Phase I, II and III. The chance it closes is 0.62 + 0.36 × climate, where the market sets climate to 0.30 (bad), 0.60 (normal) or 0.85 (good) — so a raise succeeds about 73% of the time in a bad market and about 93% in a good one, per attempt, compounded across three gates. The three settings bracket the observed range, from the 2022–23 winter to a 2021-style peak. Phase III is scaled slightly harder than Phase I (×1.02 vs ×0.95) because a pivotal round is the largest. Two adjustments then apply: a program already past 55% of its benchmark clock loses a further 15% (×0.85), reflecting investor fatigue with a program that is dragging; and a program that has compressed its time and cost needs a smaller, shorter round, which scales its chance of failing to raise by 0.15 + 0.85 × need (need = this round's time-and-cost against benchmark, so exactly neutral at benchmark and easier below it). These coefficients are calibrated to land the pooled likelihood of approval and the share dying of financing in a plausible range — they are not drawn from a specific dataset, and this gate is the weakest structural assumption in the model.
Modelled vs data-anchored. The financing-gate switch replaces all of that with a curve tied to real numbers. Biotech follow-on runs about 50% per round in normal markets; in the 2022–23 winter SVB counted 356 Series A biotechs producing only 102 Series B (~29%), and EY put 55% of emerging biotechs at under two years of cash. The data-anchored curve is calibrated to that spread (~4% fail-to-raise in a good market, ~23% in a bad one) and drops the drag and round-size add-ons. Honest limits: observed graduation conflates money with science and M&A, so the viable-but-unfunded share can't be isolated cleanly; funding rounds don't map one-to-one onto clinical phases; the winter figure is one lender's book. The two modes land close, which is some reassurance — but data-anchored removes the "faster programmes raise more easily" effect, because that isn't visible in the data.
The rest. Costs are out-of-pocket and include company overhead and the capital burned by companies that died of financing rather than science. The China route charges no extra time or money for a US bridging study, only the Phase I speed and cost gains, so it reads a little favourably. The time switches use vendor-reported reductions and stack multiplicatively; nothing audits them. Failure-mode splits come from older, coarse literature. The calendar years start from the financing climate you pick, but that climate is then held fixed for the whole run — a 2021 cohort does not live through 2022's crash here, which would mean treating the start year as a vintage rather than a setting.
Each diagram animates one cohort of 100; the comparison table pools 60 cohorts per scenario (6,000 programmes each). Both scenarios draw from the same fixed random strips, so the table's differences are attributable to the switches — but the levels themselves still carry sampling error of roughly ±8%.
The reason a surviving drug costs what it does is visible the moment you switch the model from measuring companies to measuring capital. Killing a program in pre-clinical costs almost nothing. Killing one in Phase III, after years of trials and hundreds of patients, is where the money actually goes — and that cost of failure, spread across the whole cohort, is what the survivor has to carry. The written-off bar reddens toward the right precisely because the losses that land late are the expensive ones.
The model is also a tool for asking how AI and other changes might move those numbers. You can run two scenarios side by side against the identical cohort — same companies, same luck — and toggle interventions grouped into two kinds: those that save time and money (generative chemistry, AI-run trial operations, running Phase I in China) and those that improve the actual odds of success (in-silico toxicity prediction, genetically supported target selection). Because both scenarios share their random draws, any difference between them is caused by the switches rather than by chance.
The recurring lesson is that levers which merely go faster help far less than levers that change whether a drug works, because cost-per-approval is driven by the probability of success. The single best-evidenced intervention in the model — choosing targets with human genetic support — outperforms every speed-and-cost improvement combined. Every number is sourced from published benchmarks (Citeline/BIO, Tufts CSDD, DiMasi, Nelson/Minikel), with the assumptions and their weaknesses stated openly at the bottom of the model.